PodFailPred: A New Open-Source Dataset to Train AI Models on Predicting Pod Failures in Kubernetes

The growing reliance on microservices architecture has introduced new challenges in maintaining system reliability, particularly in predicting the failure of microservices workloads.Previous research has highlighted the potential of using artificial intelligence for this purpose but has been limited by the absence of dedicated and detailed datasets.This paper introduces PodFailPred, a dataset generated by the use of chaos engineering to capture interdependencies between infrastructure, network, and application-layer failures in Kubernetes-based microservices environments.Our rigorous methodology systematically injects and monitors 15 distinct failure types.The resulting dataset encompasses 21,374 entries with 38 distinct metrics, providing unprecedented coverage of pod lifecycle events and failure patterns.Evaluation demonstrates the dataset's effectiveness through machine learning models achieving 99.94% accuracy in failure prediction, significantly outperforming existing approaches.PodFailPred is publicly available, offering a robust foundation for training AI models to develop reliable failure prediction systems.